Research And Implementation On Detection Algorithm For Abnormal Group Behavior Based On Global Optical Flow

In the field of computer vision, the computation of optical flow and the research of three-dimensional reconstruction technology based on image sequences have become a hot issue. Since1981, Horn and Schunck proposed the classic optical flow algorithm, which provided the reference of optical flow field research for many experts and scholars at home and abroad. At the same time, optical flow algorithm is playing a more and more important position in all fields of social life, including industrial robot vision system, unmanned aerial vehicle target detection and navigation system, automatic analysis of space satellite photos and tracking system, analysis medical image and diagnosis system, etc, especially, the applications of image segmentation and moving target detection are more outstanding, at present, it is widely used in medical field, military field, intelligence field, etc.In view of the complex scenario, the Horn-Schunck global optical flow algorithm is studied and improved in the paper and detection algorithm of abnormal group behavior based on the global optical flow is proposed, specific research steps and innovation points of this paper are as follows:Corresponding to the global smoothness constraints of Horn-Schunck global optical flow algorithm, which forces the estimated optical flow smoothly through each area, when the selection of smoothing term coefficient factor is too large, it will smooth out the important information about the shape of object, and bring greater error of optical flow vectors. This paper processes image based on total variation regularization (ROF) model, and then get texture image, which works as the input image for optical flow calculation.After processing the input image by a Gaussian filter, using multi-resolution hierarchical refinement method to get image pyramid structure which does not need to consider the limit of large displacement movement of the original image, reducing the amount of displacement in the image layering process. Initializing each layer optical flow vector by bicubic interpolation, and then making optical flow solution reach steady state through finite iteration, due to the abnormal optical flow vector caused by noise and other factors, the adaptive neighborhood correction method to correct abnormal optical flow vector is proposed to improve the accuracy of algorithm.The group target is considered as a whole, each pixel of input video image is regarded as a particle, use the positive and reverse combination way to calculate the Finite time Lyapunov exponent, and then perform a scene classification. Particle attributes set is utilized to character the behavior of particle, which includes instantaneous dynamic energy function and the global energy function:the instantaneous dynamic energy function has microscopic characteristics; the global energy function, obtained based on the equivalence of Markov random field and Gibbs random field, has global characteristics. On this basis, the abnormal behavior is detected.